Slow your scroll

Quick Take · Lead · Put AI to Work · Looking Ahead

In this issue

The "30 million gallons of water in secret" story everyone shared this week — and why progressive misinformation about AI keeps costing us
Trump's executive order on frontier AI model vetting could drop in the next two weeks. It's both more interesting and more complicated than it looks
Recidiviz is using AI to help people get out of prison faster, not to keep them in. 19 states, 250,000+ people accelerated out of the system
A five-minute AI workflow for fact-checking the next viral "AI is destroying X" headline before you share it

Quick Take

There's something specifically progressive about the way we read AI stories right now, and most weeks that's an advantage — we notice the labor angles and accountability gaps that other readers miss. But the inflammatory anti-AI story is traveling so fast in progressive spaces that we're starting to share things that don't quite hold up. Every time we do, we hand the win to people who'd love for progressives to be unreliable on AI.

This issue is about getting that right. One story this week traveled like wildfire, and the local paper's coverage four days later told a different story. Another is barely traveling at all, and it's the one that's actually going to shape AI policy for the next four years. Critical reading, both directions.

Let's get into it.

4 days

The gap between the viral national story about QTS Fayetteville (E&E News, May 7) and the local paper's corrective version (The Citizen, May 11). Four days for the most accurate account to surface, and most of us had already shared the first one. Worth waiting for.

Source: The Citizen

The 30 million gallon story everyone shared this week

The story trending in progressive feeds this week: an AI data center used 30 million gallons of water "in secret" while a Georgia county asked residents to conserve. It traveled because it sounded like exactly what we already think is happening: Big Tech taking what it wants while communities hold the bag.

The version that traveled isn't quite the version that's in the record.

QTS, a Blackstone-owned data center developer, is building a 615-acre campus in Fayetteville, Georgia. Roughly 8,000 construction workers on site at peak. The water in question wasn't running AI servers. It was being used to pour concrete and keep all those workers' bathrooms running. While that work was happening, Fayette County was migrating 33,000 water meters from an old system to a new smart-meter platform. A QTS meter didn't make the transition successfully, and a second connection wasn't linked to QTS's billing account. About 29 million gallons flowed before anyone caught the gap.

When the county figured it out a year ago, they sent QTS a bill for $147,474. QTS paid. The letter sat in county files until earlier this month, when a local attorney pulled it through an open records request and posted it to Facebook. An advocacy group surfaced it, Politico ran the national piece, and Tom's Hardware put the word "secretly" in their headline.

That word didn't appear in the primary reporting. It also doesn't match what the county administrator told the local paper this week. He said staff had physically inspected the meter during installation and had simply lost track of it during the system upgrade, and acknowledged the letter's wording made the situation sound more sinister than it was.

That's the boring version. None of it is comforting, but none of it is a stealth water heist either.

Why this matters. There's a real story under this one. It's just not the heist story. The county's water director, asked why QTS wasn't fined, told reporters:

"They're our largest customer, and we have to be partners. It's called customer service."

That's the line that should make progressives twitch. Not the meter. The reflex that a private-equity-owned industrial water customer gets "customer service" while the rest of the county is on a Level 1 Drought Response and being asked to stop watering their lawns. Fayetteville's City Council already saw enough. They banned new data centers in every zoning district earlier this year. The policy story is real, and it's already moving locally.

Here's why we care about this in this newsletter. The version that traveled wasn't the policy story. It was the heist version. Every time we share an inflammatory frame that doesn't quite hold up, somebody with an afternoon and a substack pulls the county administrator's quote, points out the water was for concrete, and we look like the people who didn't read past the headline. The pro-AI-no-questions side then gets to use us as the example of why progressive concerns about AI are unreliable. We hand them the win on a story we should have won on the merits.

We can be the people who get this right. The accurate version of the QTS story is more useful to the progressive case than the viral one, because it's defensible. "A small county utility had no infrastructure to monitor an industrial customer drawing millions of gallons during drought, and chose not to penalize them when they found out" is a real policy gap. It's the kind of thing you can take to a state legislator. "They did it on purpose" gets you a viral repost and a fact-check that follows you for the next six months.

What you can do

Slow your scroll on anti-AI headlines. There's real momentum building right now to blow every AI story out of proportion, and the inflammatory version is almost never the one that survives a second look. Before you share the next "AI is destroying X" piece, spend 90 seconds checking it. Look for the local paper's coverage instead of the aggregator's, and check whether the eye-popping number applies to one specific facility or got scaled up to "all of AI forever." If the only thing behind the alleged harm is an outraged tweet, hold off. We can be critical of AI without being credulous about every story that confirms what we already think.

The other AI story this week

While the QTS water story was eating progressive feeds, Kevin Hassett went on Fox Business on May 6 and said the White House is studying an executive order to make pre-deployment review of frontier AI models mandatory. He compared it to FDA drug approval and said signing could come within two weeks. The day before, the Center for AI Standards and Innovation — the rebranded, partly-hollowed-out version of Biden's AI Safety Institute — announced voluntary review agreements with all five major US frontier labs: OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI. On Monday, Americans for Responsible Innovation published a proposal to tie that safety review to federal contract eligibility.

The trigger isn't abstract. Anthropic recently published a red-team report on a model it calls Mythos and chose not to release it. Mythos can autonomously find and chain zero-day vulnerabilities across every major operating system and browser. Anthropic's own report documents the model finding thousands of critical vulnerabilities at under $20,000 per thousand runs, and chaining four of them together to escape browser and OS sandboxes. The NSA was already testing it to find flaws in Microsoft products before the new CAISI agreements were announced.

That's the kind of capability that makes "should the government look at these models before they ship?" a genuine question, not an academic one.

Why this matters. The progressive instinct on federal AI accountability has been "yes, please." A pre-deployment review regime is, on paper, exactly the kind of accountability we've been asking for.

The complication is who's building it.

CAISI, the agency doing the vetting, is the same Biden-era safety institute the Trump administration partly hollowed out. The AI Action Plan stripped DEI, climate, and misinformation references from NIST's risk framework. And in December 2025, Trump signed Executive Order 14365, which preempts state AI laws covering the broader harms progressives have spent years legislating against: hiring discrimination, labor protections, election integrity. The Center for American Progress called that order an unprecedented threat to state-level AI protection.

So here's the shape of what's being built: a federal review regime focused on national-security-class harms, designed and run by the same administration that just dismantled the state-level protections covering everything else. Hassett's FDA analogy doesn't quite hold either. FDA can block drugs from going to market. The proposed AI regime, as reported so far, can review models but not block them. That's disclosure theater, not gatekeeping.

The progressive read isn't "oppose any federal review." A safety floor is better than no safety floor, and the Mythos case shows there are real reasons to want one. The read is: if the EO ships with no blocking authority, stays narrow on scope, and leaves state preemption in place, we're being shown a thin federal accountability regime as a substitute for the actual one we just lost.

What you can do

Watch the next two weeks. Hassett floated a signing window of roughly May 13 to May 20. Three things to look for when an EO actually drops. One: does it give CAISI authority to delay or block a release, or only to evaluate? Two: is the scope national-security-only, or does it cover bias, labor, and election harms that state laws used to cover? Three: does it touch the state-preemption order from December, or leave that intact? Narrow scope plus no blocking authority plus continued state preemption isn't accountability. It's the appearance of accountability while the real ones stay disabled. That's the story to tell clearly when it happens.

Progressive AI Win

Recidiviz uses AI to get people out of prison faster — not to keep them in

The usual AI-in-criminal-justice story is the one progressives have spent a decade fighting: predictive policing and risk-score algorithms aimed at defendants. Recidiviz, led by CEO Clementine Jacoby (a former Google PM and TED Fellow), is the opposite version. Their tablet app puts information directly into the hands of roughly 40,000 incarcerated people: personalized sentence dates, earned-time tracking, and program eligibility that residents previously had to ask staff for. On the staff side, their AI Case Planning Assistant uses fine-tuned models to cut about an hour off every reentry assessment, so case managers can move more people through reentry prep, faster. In April, Missouri became their third sentencing-partner state.

The scale is real: 19 state partnerships, about 46% of the US incarcerated population reached, an estimated 250,000 people accelerated out of the system since 2019, and roughly $1.5 billion in savings for partner agencies. Fast Company named them one of the most innovative companies of 2026 in March.

"98% of incarcerated people eventually return to their communities. Yet most don't receive meaningful reentry preparation until the very end of their sentences." — Recidiviz

This is what pro-AI, pro-accountability looks like in practice. AI used for people in the system — to speed their exit — not against them.

Put AI to Work

Practical ways progressives can use AI this week

Fact-check a viral AI story in five minutes

This issue's tagline, critical reading both directions, is also a habit you can build. You don't need a journalism degree to do what reporters already do: trace a story back to its source. AI tools make this fast. Here's a five-minute workflow for the next "AI is destroying X" piece that lands in your feed.

Step 1. Find the primary source. Aggregators rewrite other people's reporting. Find the original.

Paste the headline into Claude or ChatGPT with this prompt:

"Find the primary news article this headline is based on. I want the original outlet, not an aggregator, and I want the byline, publication date, and a direct URL."

For this week's QTS story, that prompt leads you to Politico's E&E News piece, not Tom's Hardware. Same underlying event, very different framing.

Step 2. Check whether a local paper covered it. National outlets summarize. Local reporters call the people involved. The most accurate version of a regional story is almost always in the regional paper.

"Is there local newspaper coverage of [event] from a paper in [city/county]? Look for follow-up reporting that includes direct interviews with local officials."

The Fayette County Citizen ran their piece four days after the viral version, with on-record quotes from the county administrator. It corrected the framing entirely.

Step 3. Pressure-test the eye-popping number. When a stat sounds too clean to be real, get scale.

"Put [the number in the headline] into context. What's it equivalent to for an average household? A city's daily supply?"

The number often sits in a less alarming range than the headline implied. Sometimes it sits in a more alarming one. Either way you'll know.

Step 4. Verify the dramatic quote. Headlines often turn paraphrases into apparent quotes. If an article says someone "admitted" or "conceded," check the original.

"In [original article URL], find the exact direct quote from [official name] about [topic]. Give me the wording verbatim, not a paraphrase."

Tool notes. Claude and ChatGPT are good for synthesis. Perplexity is faster when you want inline citations on every claim. For anything you need to verify yourself, ask for URLs and click through. AI is a research multiplier, not a replacement for reading. The whole point of this exercise is to make sure the source got read by somebody.

From our friends

Your org deserves its own AI. Not Big Tech's.

Change Agent is a private AI platform built for nonprofits, unions, and advocacy orgs. Your data stays yours, it plugs into tools you already use (Google Drive, Slack, ActBlue), and it handles the tedious stuff so your team can focus on the mission. Starts at $35/month. Small nonprofits under $1M can apply for discounted pricing.

Learn more

Looking Ahead

A few things to keep an eye on this week and next.

The Trump frontier-model vetting EO could land any time in the May 13–20 window Hassett floated. When (or if) it drops, run it past the three-question test from Story 2: blocking authority or just review authority? National-security-only scope, or does it cover the broader harms? Does it touch December's state-preemption order? The answers tell you whether this is real accountability or staged accountability.

The Americans for Responsible Innovation proposal is the parallel track, tying safety review to federal contract eligibility for any lab spending $100M+/year on compute. Structurally sounder mechanism, more legal teeth. Watch for congressional pickup.

And the next viral AI story is already being written. Run it through the five-minute workflow before you share it.